def launch_cmaes_pure(center,
sigma,
nbeval=10000,
display=True,
ma_func=myEvalFn.sphere):
res = purecma.fmin(ma_func, center, sigma, maxfevals=nbeval)
return res[1].result[1]
def launch_cmaes_pure_nn(center, sigma, nbeval=10000, display=True, ma_func=my_Function):
res = 0
aux = center
nbiter = 0
unit_eval = 100
while(res != -500)and nbiter < nbeval/unit_eval:
print(-res)
aux,cma = purecma.fmin(ma_func,aux,sigma,maxfevals=unit_eval,verb_disp=0)
res = cma.best.f
nbiter += 1
return -res,aux
def compare(self, func, mu_0, min_f, choice_prior, verbose):
tol1 = 1e-30
tol2 = 1e-30
dim = len(mu_0)
mu_0 = np.array(mu_0)
n = 4 + int(3 * log(dim))
# mu_0 = np.array([400,400])
k_0 = 4
v_0 = n + 2
factor = 1
psi = np.eye(dim)
list_param = [mu_0, k_0, v_0, psi, factor]
np.random.seed(3)
list_x_star_1, val1 = self.compute_min(list_param, n, tol1, tol2, func,
'StrategyOne', choice_prior,
verbose)
np.random.seed(3)
list_x_star_2, val2 = self.compute_min(list_param, n, tol1, tol2, func,
'StrategyTwo', choice_prior,
verbose)
random.seed(3)
res = pcma.fmin(func, mu_0, 1, verb_disp=1000)
data = res[1].logger._data
cma_esvalues = [f for f in data['fit']]
cma_esvalues = cma_esvalues[1:]
best_seen = cma_esvalues[0]
cma_bestvalues = []
for elem in cma_esvalues:
if elem < best_seen:
cma_bestvalues.append(elem)
best_seen = elem
else:
cma_bestvalues.append(best_seen)
val1 = np.abs([i - min_f for i in val1])
val2 = np.abs([i - min_f for i in val2])
cma_bestvalues = np.abs([i - min_f for i in cma_bestvalues])
plt.figure(figsize=(6, 4.5))
plt.semilogy(np.arange(len(val2)), val2, label='B-CMA-ES S2')
plt.semilogy(val1, label='B-CMA-ES S1')
plt.semilogy(cma_bestvalues, label='CMAES')
plt.title(func.__name__ + ' convergence using ' + choice_prior +
' prior')
plt.xlabel('iteration steps')
plt.ylabel('error')
plt.legend()
plt.savefig(self.path_plot + '\\convergence_' + func.__name__ + '_' +
choice_prior + '.png')
plt.show()
[(x) for x in zip(data_test_s, label_test)])
# ------Readout---------------
states_train, states_test, _label_train, _label_test = [], [], [], []
for train in states_train_list:
states_train.append(train[0])
_label_train.append(train[1])
for test in states_test_list:
states_test.append(test[0])
_label_test.append(test[1])
states_train = (MinMaxScaler().fit_transform(np.asarray(states_train))).T
states_test = (MinMaxScaler().fit_transform(np.asarray(states_test))).T
score_train, score_test = readout.readout_sk(states_train,
states_test,
np.asarray(_label_train),
np.asarray(_label_test),
solver="lbfgs",
multi_class="multinomial")
# ----------show results-----------
print('parameters %s' % parameter)
print('Train score: ', score_train)
print('Test score: ', score_test)
return 1 - score_test
##########################################
# -------CMA-ES parameters search---------------
if __name__ == '__main__':
core = 10
pool = Pool(core)
res = purecma.fmin(parameters_search, [0.5, 0.5, 0.5], 1, verb_disp=100)
def launch_cmaes_pure(center, sigma, nbeval=10000, display=True, ma_func=sphere): aux = purecma.fmin(ma_func,center,sigma,maxfevals=nbeval)[0] return ma_func(aux)
from brian2 import *
from multiprocessing import Pool
from cma import purecma
def aaa(inputs):
N = 1
print(inputs, id(N))
return inputs + 1
def parameters_search(parameter):
# ---- check parameters >0 -----
if (np.array(parameter) < 0).any():
return np.random.randint(10, 100)
states_train_list = pool.map(aaa, [1, 2, 3, 4, 5])
print(states_train_list)
return np.random.randint(5, 100)
##########################################
# -------CMA-ES parameters search---------------
if __name__ == '__main__':
core = 2
pool = Pool(core)
res = purecma.fmin(parameters_search, [30, 1, 1], 1, verb_disp=100)
def compare(func, mu_0, min_f, nb_seed, save_fig):
multivariate = True
tol = 0.005
# CMAES standard population size
n = 4 + int(3 * math.log(len(mu_0)))
k_0 = 4
v_0 = n + 2
factor = 1
psi = np.eye(2)
list_param = [mu_0, k_0, v_0, psi, factor]
Function_evaluations = np.zeros(0)
bcma_v = np.zeros(0)
cma_v = np.zeros(0)
iteration_max = 500
for s in range(nb_seed):
np.random.seed(s * 3)
list_x_star, values = compute_min(list_param, n, tol, multivariate,
func)
bcma_bestvalues = []
for i in range(0, len(values), n):
bcma_bestvalues.append(values[i] - min_f)
if i < len(values) - 1:
bcma_bestvalues.append(values[-1] - min_f)
random.seed(s)
res = pcma.fmin(func, [mu_0[0], mu_0[1]], 1, verb_disp=0)
data = res[1].logger._data
cma_esvalues = [f for f in data['fit']]
cma_esvalues = cma_esvalues[1:]
best_seen = cma_esvalues[0]
cma_bestvalues = []
for elem in cma_esvalues:
if elem < best_seen:
cma_bestvalues.append(elem)
best_seen = elem
else:
cma_bestvalues.append(best_seen)
cma_bestvalues = np.abs([i - min_f for i in cma_bestvalues])
max_length = max(min(len(bcma_bestvalues), len(cma_bestvalues)), 30)
for i in range(max_length):
Function_evaluations = np.append(Function_evaluations, i)
cma_v = np.append(cma_v, cma_bestvalues[i] \
if i < len(cma_bestvalues) else cma_v[-1])
bcma_v = np.append(bcma_v, bcma_bestvalues[i] \
if i < len(bcma_bestvalues) else bcma_v[-1])
if max_length < iteration_max: iteration_max = max_length
df1 = pd.DataFrame()
df2 = pd.DataFrame()
df1['Function_evaluations'] = Function_evaluations
df1['Method'] = ['CMA-ES'] * len(Function_evaluations)
df2['Function_evaluations'] = Function_evaluations
df2['Method'] = ['BCMA-ES'] * len(Function_evaluations)
df1['Error'] = cma_v
df2['Error'] = bcma_v
df = pd.concat([df1, df2], axis=0)
df = df[df['Function_evaluations'] <= iteration_max]
f, ax = plt.subplots(figsize=(8, 5.2))
ax.set(yscale="log")
if not save_fig:
plt.title('Convergence comparison between CMA-ES and BCMA-ES\n' +
func.__name__)
sns.lineplot(x="Function_evaluations", y="Error", data=df, hue="Method")
if save_fig:
print(f'saved figure {func.__name__}_convergence.png')
plt.savefig(f'{func.__name__}_convergence.png')
plt.show()
return df
from brian2 import *
from multiprocessing import Pool
from cma import purecma
def aaa(inputs):
N = 1
print(inputs, id(N))
return inputs+1
def parameters_search(parameter):
# ---- check parameters >0 -----
if (np.array(parameter)<0).any():
return np.random.randint(10,100)
states_train_list = pool.map(aaa, [1,2,3,4,5])
print(states_train_list)
return np.random.randint(5,100)
##########################################
# -------CMA-ES parameters search---------------
if __name__ == '__main__':
core = 2
pool = Pool(core)
res = purecma.fmin(parameters_search, [30,1,1], 1, verb_disp=100)
def doPcma():
gaussHolder.reset()
gaussHolder.indexLabel.configure(text="Gaussian 1")
angles, _ = pcma.fmin(pc.operate, initialAngles, 0.1)
arm.setAngles(angles)
arm.draw()